Large Model Training Scientist / Engineer

Large Model Training Scientist / Engineer (Pre-training)

Role Overview

You will work on the research and engineering of scientific foundation models, including model architecture exploration, training stability, and large-scale training optimization. Your work will become part of the Scinetics AI for Science R&D stack, driving continuous iteration across data, training, and evaluation and producing models that work in real scientific settings.

Research Areas

• Large Language Model Pre-training

• Vision-Language Model Training

• Multimodal Foundation Models

• Diffusion Model Training

• Model Architecture Exploration

• Distributed Large-scale Training

• Training Performance Optimization

• AI for Science

Responsibilities

• Model Training & Architecture: Own end-to-end pre-training for scientific foundation models. Explore and validate architectures for scientific use cases, including unified multimodal architectures. Build training pipelines, diagnose instability, optimize performance, and continuously improve foundation-model capabilities.

• Data & Evaluation: Build, synthesize, clean, and curate high-quality scientific data. Develop multidimensional evaluation suites for scientific tasks and use evaluation results to drive model iteration.

• Large-scale Training: Build and optimize distributed training systems to improve GPU utilization, throughput, stability, and experiment velocity.

• Research Exploration: Stay current with advances in foundation models, multimodal models, unified architectures, and related areas, and rapidly turn new research ideas into testable model experiments.

What We Look For

• Bachelor’s degree or above in computer science, artificial intelligence, mathematics, or a related STEM field.

• Strong foundations in deep learning and large models, with hands-on experience in LLM / VLM / diffusion pre-training and verifiable results such as metrics, reports, papers, or open-source projects.

• Proficiency in PyTorch and experience with large-scale distributed training frameworks such as DeepSpeed and Megatron-LM.

• Familiarity with unified multimodal / unified architectures; experience studying, reproducing, or modifying models such as BAGEL is a plus.

• Long-term interest in AI for Science and the ability to independently drive training tasks and model experiments.

Nice to Have

• Research or engineering experience in model architecture innovation or unified multimodal architectures.

• Experience with large-scale training stability, throughput optimization, memory optimization, or failure diagnosis.

• Experience building evaluation benchmarks for scientific tasks.

• Publications at strong conferences or journals.

• High-quality open-source contributions.

• Awards in AI-related competitions.